Frontier labs will increasingly keep their most capable models secret to self-advance

Frontier labs will increasingly keep their most capable models secret to self-advance | We still don't have the model that they used to win the gold medal in the IMO... that's the first bifocation that you see. We used to have the frontier model every single time. The moment they got to that, that was the last time.

Watch events: ARC-AGI-2 scores; Frontier Math Tier 4 benchmark; SWE-bench Verified; Humanity's Last Exam

We still don't have the model that they used to win the gold medal in the IMO... that's the first bifocation that you see. We used to have the frontier model every single time. The moment they got to that, that was the last time.

Beneficiaries (14)

Adverse (7)

{
  "nia": false,
  "url": "https://www.youtube.com/watch?v=d__HRChE2ZE",
  "mode": "THESIS",
  "role": "Host",
  "context": "are you going to keep your model secret because you're going to be able to use them to advance your company far faster than anybody else? He said, 'No, no. Our go our job is to get out there to the public.' I don't believe that. We still don't have the model that they used to win the gold medal in the IMO... that's the first bifocation that you see.",
  "verbatim": "We still don't have the model that they used to win the gold medal in the IMO... that's the first bifocation that you see. We used to have the frontier model every single time. The moment they got to that, that was the last time.",
  "conv_cues": "first bifurcation; I don't believe that",
  "direction": "HAPPEN",
  "timeframe": "Ongoing",
  "conv_level": "HIGH",
  "milestones": [
    {
      "kind": "llm_pre_event",
      "label": "OpenAI's IMO-gold-medal model remains unreleased months after milestone",
      "notes": "HIT - 'Hidden AI Frontier' coverage explicitly confirms OpenAI's IMO-gold model 'will not be released for months' - direct corroboration of Diamandis' bifurcation thesis.",
      "source": "https://ai-frontiers.org/articles/the-hidden-ai-frontier",
      "status": "hit",
      "weight": 0.4,
      "ordinal": -5,
      "source_id": null,
      "confidence": 0.95,
      "source_url": "https://ai-frontiers.org/articles/the-hidden-ai-frontier",
      "expected_date": "2026-04-30",
      "observed_date": "2026-04-30",
      "research_origin": "deep_research",
      "measurement_criterion": "OpenAI publicly confirms the math-reasoning model that achieved IMO gold-medal performance has not been released to the public, months after the achievement was reported"
    },
    {
      "kind": "quartile_checkpoint",
      "label": "Q1 window check-in (25%)",
      "status": "pending",
      "weight": 0.05,
      "ordinal": -4,
      "source_id": null,
      "expected_date": "2026-07-28",
      "observed_date": null
    },
    {
      "kind": "quartile_checkpoint",
      "label": "Q2 window check-in (50%)",
      "status": "pending",
      "weight": 0.05,
      "ordinal": -3,
      "source_id": null,
      "expected_date": "2026-10-25",
      "observed_date": null
    },
    {
      "kind": "quartile_checkpoint",
      "label": "Q3 window check-in (75%)",
      "status": "pending",
      "weight": 0.05,
      "ordinal": -2,
      "source_id": null,
      "expected_date": "2027-01-22",
      "observed_date": null
    },
    {
      "kind": "llm_pre_event",
      "label": "Frontier lab publicly admits internal-deployment-only model with significant capability gap vs released models",
      "source": "https://arxiv.org/html/2604.23065 (Internal Model Deployment paper); https://metr.org/common-elements",
      "status": "pending",
      "weight": 0.4,
      "ordinal": -1,
      "source_id": null,
      "confidence": 0.85,
      "source_url": "https://arxiv.org/html/2604.23065",
      "expected_date": "2027-01-30",
      "research_origin": "deep_research",
      "expected_date_range": {
        "to": "2027-09-30",
        "from": "2026-06-01"
      },
      "measurement_criterion": "OpenAI, Anthropic, Google DeepMind, or xAI publishes a system card / safety policy that explicitly references an internal-only model with capabilities 'meaningfully ahead' of public releases"
    },
    {
      "kind": "event",
      "label": "Frontier labs will increasingly keep their most capable models secret to self-advance",
      "status": "pending",
      "weight": 1,
      "ordinal": 0,
      "source_id": "238_013",
      "expected_date": "2027-04-22",
      "observed_date": null
    },
    {
      "kind": "llm_pre_event",
      "label": "AI R&D acceleration measurable: lab discloses >=20% productivity gain from internal model use",
      "notes": "Diamandis' explicit claim was that labs would use these internally to advance themselves faster.",
      "source
... (truncated)